I

Chapter 2 22 Levels of SYntactic Analysis 23

tree), which is the canonical expression of themes and patients in English.

(2.12) Elmer gave Hortense a porcupine.

According to this account, a porcupine in (2.12) is an indirect argument of the verb gave and should display precisely the same syntactic be­havior it would if it were marked by a preposition.

It seems that, in the unmarked case, each semantic role assigner may assign only one semantic role. To derive some results in chapters 5 and 7 concerning double object constructions and applied verbs, I need to elevate this observation (0 the status of a principle:

(2.13)

The One Role/Role Assigner Principle: In the unmarked case, a semantic role assigner may assign only one role.

Moreover. an argument-taking item tends to assign a role to its argu­ment, if it can. I incorporate this apparent tendency into principle (2.14).

(2.14) The Direct Argument Principle: In the unmarked case, an argument-taking item will assign a semantic role to its argument.

Principle (2.14) implies that unmarked arguments are direct arguments. The One Role/Role Assigner and Direct Argument principles are re­

strictive principles formulated by noting observable tendencies in the world's languages. In contrast. principle (2.15) is a stipulative con­straint essential to the present theory.

(2.15) The Lexical Exception Principle: Exceptions (to markedness principles) must be encoded in the lexical entries of lexical items.

Since predicates are phrases, not lexical items, they lack lexical en­tries, which might contain exceptional features. Thus, by the One Role/Role Assigner ami Lexical Exception principles, predicates will assign at must one semantic role each, and by the Direct Argument and Lexical Exception principles, the argument of a predicate will receive

its semantic role from the predicate; that is, the subject of the predicate will be a direct argument.

The notions of semantic role assignment and argument structure pro­vide definitions for the two basic semantic relations of l-s structure; that between an argumenHaking item (like a verb or predicate) and its ar­guments, and that between a semantic role assigner and the constituent to which ir assigns a semantic role. Other semantic relations may be defined in terms of these basic relations and the constituent types of I-s structure. For example, I define the logical object ofa verb as the con­stituent that both serves as argument to and receives a semantic role from the verb. The logical subject of a predicate is the constituent serving as argument and semantic role assignee of the predicate.

2.1.1.2. J-s Subject-Object Asymmetries To repeat, constituents are assigned their semantic roles in the present theory; they do not receive roles by virtue of occupying slots in predicates. Another important difference between P-A structures of this book and the sort of predicate notation employed by Jackendoff and others is that the former but not the latter imply an asymmetry among the inherent semantic roles asso­ciated with a verb. In the Jackendoff (1976) framework and in many current theories all semantic dependents of a verb simply hll slots in n-ary predicates such as GO (x, y, z). This equality of arguments char­acterizes Bresnan's Lexical-Functional Grammar, Williams's argument structures (Williams 1981), and Montague Grammar (Dowty 1982a. 1982b) as well. [0 a theory with symmetric argument structures, the choice of one argument can in no way affect the semantic role assigned to another argument of the predicate. All arguments are independent and on par. In contrast to the symmetric argument structures of these other frameworks, the asymmetric P~A structures incorporated into the present theory assure that the choice of arguments to fill P-A slots will affect the semantic role assigned to the (logical) subject. Since the P-A structures of verbs are functions from arguments to predicates, which arguments one inserts into the P-A structures determines what predi­cate the function wilt yield. As the predicate assigns the logical subject its semantic rote, choice of arguments for P-A structures determines the semantic role of the logical subject.

By way of Illustration. consider the sentences in (2.16al and (2.17a). The verb give provided with the argument' in (2.100) will yield the predicate in (2.16b); give provided with the arguments in (2.17a) will yield the predicate in (2I7b).

24Chapter 2

(2.16) a. Elmer gave the porcupine to Hortense. b. 'give' (the porcupine, Hortense) = (x give the porcupine to Hor­

tense)

(2.17) a. Elmer gave two aardvarks to Horace. b. 'give' (two aardvarks, Horace) = [x give two aardvarkli to Horace)

Since the different choice of arguments as theme and goal in (2.16a) and (2. I7a) produce different predicates, Elmer the logical subject of both sentences, will bear a distinct role in each. the t· giver of the porcupine to Hortense" role in (2.16a) and the "giver of two aardvarks to Horace" role in (2.17a). But regardless of what NP is chosen as subject of give the porcupine to Hortense. the porcupine will be assigned the

same theme role by give. The proposed distinction between the logical subject and the verb's

arguments implies an asymmetry between the effects of changing the arguments of a verb and the effects of changing the logical subject of a predicate. Choice of arguments for a verb can affect the semantic role assigned to the logical subject, but in no way may the choice of Logical subject affect the semantic rules assigned to the verb's arguments. Al­though the predictions of this asymmetric treatment of a verb's seman­tic dependents are clear. they are not easy to test. Compare sentences

(2.16a) and (2. I7a) with sentences (2.18).

(2.18) a. Elmer sold two porcupines to Hortense. b. Horace sold two porcupines to Hortense.

1 claimed without argument that Elmer bears the "giver of the por­cupine to Hortense" role in (2.160) but the "giver of two aardvarks to Horace" role in (2.17a)-two clearly distinct roles assigned to Elmer by the distinct predicates that the' give' function produces when handed distinct sets of arguments. One could with equal justification claim that two porcupines bears the "sold by Elmer" role in (2.18.) but the "sold by Horace" role in (2.18b) instead of bearing simply the sold or theme role in both sentences. Clearly, being sold by Elmer differs from being sold by Horace. What is not clear is whether this differ­ence demonstrates that two porcupines is assigned different roles in

(2. j Sa.b).

Levels of Syntactic Analysis 25

If 1 restricted attention to examples like (2.16)-(2.18), 1 could not argue for the proposed asymmetry among verb dependents. In such examples, changing the logical subject seems to have as much effect on the semantic role borne by the verb's logical object as changing the logical object has on the semantic role borne by the verb's logical sub­ject, and the effect in both cases is minimal. Expanding the data base, however, yields convincing evidence in English for the predicted asymmetry. Just about every simple transitive English verb expresses a wide range of predicates depending on the choice of direct object. The predicates in (2.19) illustrate this phenomenon.

(2.19) a. throw a baseball b. throw support behind a candidate c. throw a boxing match (i.e., take a dive) d. throw a party e. throw a fit f. take a book from the shelf g. take a bus to New York h. take a nap i. take an aspirin for a cold J. take a Jetter in shorthand k. kill a cockroach I. kill a conversation m. kill an evening watching T. V. n. kill a bottle (i.e., empty it) o. kill an audience (i.e.. wow them)

One might distinguish the basic or literal uses of the verbs in (2.19) from their metaphoric and idiomatic uses, yet no clear lines divide (he classes. Although the different predicates formed by adding different direct objects to most simple transitive verbs assign different semantic roles to their subjects, one would be hard pressed to argue that a differ­ent predicate implies a different (homophonous) verb in each case. The person throwing a baseball propels it through the air with a motion of his arm; the person throwing a party may do no more than telephone a few guests; the person throwing a fit thrashes about on (he floor. De­spite these differences in subject semantic role. may one not maintain that the verb throw in at least the majority of its uses names the same function from arguments to predicates. with the variety of predicates displayed in (2.19) being attributabfe to the variety of input arguments

! Chapter 2 26 Levels of Syntactic Analysis 27

to this function? Killing a cockroach involves causing it to become not alive; killing an evening may be accomplished by sitting motionless in front of aT. V.; and killing a bottle implies causing it to become empty, yet some thread of meaning ties these uses of kill together, a thread we may weave into a single function, 'kill: from arguments to predicates. Par present purposes, the importance of the range of predicates illus­trated in (2.19) lies in the dependence of the predicate on the choice of arguments for the verbs.

It might be claimed that the semantic roles borne by the direct ob­jects in (2.19) also differ from use to use of the verbs. Thus a book in (2 .19f) may be called a theme since it undergoes a change in position, whereas a letter in (2.19j) is something created, undergoing a change in state, perhaps, but not necessarily a change in position. Although the identification of the particular semantic or thematic roles assigned by verbs is a vexing problem (see section 2.l.2), regardless of whether the direct objects in (2.19) do differ in semantic role from use to use of the verbs, the choice of subject for the verbs does not determine the semantic role of their objects. Consider the sentence schemas in (2.20).

(2.20) a. The policeman threw NP. b. The boxer threw NP. c. The social director threw NP. d. Aardvarks throw NP. e. Throw NP' f. Harry killed NP. g. Everyone is always killing NP. h. The drunk refused to kill NP. i. Silence can certainly kill NP. j. Cars kill NP.

Although the more material these sentence fragments contain and the more elaborate the context one sets them in, the more they suggest particular direct objects and therefore particular predicates, the differ­ent subjects for the verbs do not imply different semantic roles for the object NPs. Note that certain subjects may violate the selectional re­strictions of certain predicates, thereby limiting the range ofpret..licates that may happily cooccur with them. Thus Silence threw NP does not readily support a reading in which throw produces the predicate it does in Harry threw tile baseball, Varying the subject of a transitive verb, however. simply does not produce a range of predicates on objects

similar to the range of predicates on subjects produced by varying the objects of transitive verbs. This is the asymmetry that confirms my conception of verb semantics; choice of object (or other argument of a verb) affects the semantic role of the logical subject whereas choice of logical subject does not affect the semantic role of the object.

The compositional 3.".ymmetry between the logical subject and a verb's arguments explains another striking property of English. New predicates in English are created by modifying a verb's function from arguments to predicates. That is, new values are given to the function for certain special input arguments. As a result of the asymmetric com­positional semantics of sentences, then, there are countless object idioms in English like kirk the bucket while subject iujom~ that are not also full phrasal idioms are rare, if they exist at all. Object idioms are combinations of verb plus object with slightly or highly unusual seman­tics from what one would expect from the canonical uses of the verb. A subject idiom would be an idiomatic combination of subject and verb. I know of only one such idiom with a free object position. Bresnan (I982a), in a criticism of an earlier version of this work, points out What's eating NP, meaning 'what's bothering NP,' as a possible exam­ple. This particular idiom has a peculiar property not always associated with object idioms: its sentential syntax is fixed. One can change the idiom to the past tense- What l1.'as eating NP yesterday?-but no sen­tence internal modifications arc allowed (it is argued in sections 2.3.3 and 3. L that tense should be considered a sentence-external modifier). For example, the idiom requires progressive aspect: What ate NP? lacks the idiomatic reading. From the point of view of the present the­ory. it is important that this apparent subject idiom has no S-internal syntax, for it is precisely S-internal syntax that is at issue, What's eat­ing NP? is not a combination of subject and verb, forming a predicate on the object, but rather a combination of wh- question syntax, pro­gressive aspect, plus subject and verb-that is. a complete sentence frame-with an open slot for an argument. Apart from the sentence frame Whm's eating x, although one does find idiomatic expressions with fixed subjects-for example. The shit hit the fan-the objects in such expressions are also fixed. One cannot say, The shit hit the air conditioner when things really go wrong.'

The evidence from the preponderance of object over subject idioms simply reinforces the conclusions derived from the range of predicates in (2.19). Predicates are created by functions on objects and other ar­guments excludmg the logical subject.

28Chapter 2

Bresnan (l982a) has objected to the claims of this section on two grounds: that my argument depends on faulty logic and that its conclu­sions are factually inaccurate. On the first point, Bresnan is simply being unfair. She attributes a conclusion to me that I do not support, then points out that my argument does not prove this conclusion.

The assumption is that if the subject is always the last argument to be semantically composed with the predicate. one can explain the gener­alization that the choice of nonsubject arguments does not depend on the choice of the subject argument. By suppressing the subject argu­ment from the verb's predicate argument structure, one prevents the subject from being combined directly with the verb before the verb and its nonsubject arguments have been assembled into a predicate; the subject can then be semantically composed only with a completely formed predicate. But the issue of whether or not the subject argument has a special role in the semantic composition of the sentence is logi­cally independent of the issue of whether or not a subject argument position should appear in lexical predicate argument structure. For ex­ample, in Dowty's theory [Dowty 1982a], the subject is always the last argument to be semantically composed with the predicate; yet the lexi­cal function that expresses the meaning of a transitive verb in his theory contains variables for both the subject and the object arguments. In short, one could capture the subjectlnonsubject generalization without affecting the lexical representation of predicate argument structure, simply by giving the subject a distinguished role as final argument in the semantic composition of the sentence. (Bresnan 1982a, 350)

Bresnan somehow takes the foregoing discussion as an argument for a particular representation of semantic subjectJnonsubject asymme­tries. It should be clear that all I am doing at present is arguing for the asymmetries themselves. The arguments for the particular representa­tion of the asymmetries within the present theory come from consid­erations of the operation of the theory as a whole. I choose the present form of representation for these asymmetries-excluding the subject from the P-A structure-because it both correctly captures the seman­tic asymmetries, as Bresnan points out in the quoted passage, and al­lows me to relate semantic subjectJnonsubject asymmetries to parallel syntactic asymmetries. It also interacts with other principles to cor­rectly constrain the possible syntactic effects of derivational morphol­ogy, as argued in chapters 4 through 7. In chapter 8 I argue specifically against Dowty's (l982a) representation of subjectlnonsubject semantic asymmetries. In summary. Bresnan is correct in pointing out that the foregoing arguments alone do not choose a representation for semantic subjectJnonsubject asymmetries, but this point is of little importance.

Levels of Syntactic Analysis 29

As Bresnan herself notes. my notation correctly predicts the asym­metries; this much is sufficient at this point in the presentation of the theory.

Turning now to Bresnan's factual quarrels with my arguments, I re­spond to each of her points in tum.

First, there are in fact subject idioms with free non-subject arguments: for example, The eat's gOIx's tongue ('x can't speak'), What's eating x? ('What is making x so irritable?'), Time's lip (for x) ('The time (for x) has expired'), x's goose is cooked ('x is in trouble and there is no way out'), (Bresnan 1982a, 350)

Note that only What's eating x? actually meets the schema for a po­tential counterexample to the claim that there are no subject idioms with free object positions. Although they contain idiomatic subject material, the other idioms Bresnan lists do not have free objects. We may dismiss x' s goose is cooked as irrelevant, as this is a passive sen­tence whose subject is presumably its logical object (cf. They cooked his goose). My arguments obviously imply that idiomatic material should not appear as a logical subject (in a sentence with a free, i.e .. nonidiomatic predicate-internal argument position); the compositional semantics of a sentence are represented directly only at I-s structure, not at s structure, where grammatical relations are encoded. This leaves The cat' s got x's tongue and Time's up (jor x). The latter example falls together with another set of examples Bresnan brings up and will be discussed with them later. The former raises some interesting ques­tions, but is not relevant to the issue at hand (I in fact mention The cat:s got x's tongue in Marantz 1981b, to which Bresnan is replying). A study of A Dictionary of American Idioms (Boatner and Gates 1975) reveals that idioms with a free possessive slot are perhaps the most common phrasal idioms with free argument positions. How one should represent the structure of these idioms is an open question. For example. in Elmer gOIher goat, what sort of argument is her? One has a free hand in the analysis of such idioms, at least currently; therefore, they cannot be used to argue against the subjectJobject asymmetries in idioms.

Bresnan continues:

Second, there are clear cases in which the semantic choice of a non­subject argument does depend upon the choice of the subject. Con sider, for example, (6a-c). (6) a. The ceiling caved in on John. b. The wall caved in on John. c. The roof caved in on John.

~

Levels of Syntactic Analysis 31Chapter 2 30

Example (6c) has a figurative or metaphorical sense ('Everything went wrong for John') as well as the literal sense that part of the structure of a house collapsed on John; but examples (6a,b) are unambiguously lit­eral. Thus, the choice of the roof as subject of caved in gives rise to a special meaning. Under this special meaning, one can choose as the object of on an abstract noun phrase which cannot occur with the literal meaning: (7) a. #The ceiling caved in on John's dreams. b. #The wall caved in on John's dreams. c. The roof caved in on John's dreams. How do we know that it is not the object of on which gives rise to the special meaning and thereby determines the choice of the subject the roof? When we omit the on phrase altogether, we find that The roof caved in still has the figurative sense 'Everything went wrong' while The ceiling caved in and The wall caved in lack it. It is not difficult to find other examples which support the same conclusion (e.g. A truck hit John vs. An ideo hit John . . .). (Bresnan 1982a, 350-351)

Bresnan's examples fail to support her point On a number of grounds. First, note that the free argument in The roofcaved in on NP is not the object of the verb. Principles yet to be described in detail make it per­fectly plausible that when a sentence lacks a syntactic object, the syn­tactic subject of the sentence may in fact be the logical object of the verb. There is good evidence that groups of verbs in many languages arc syntactically intransitive but take logical objects-the unaccusa­tive verbs of Relational Grammar. The logical objects of such verbs become syntactic subjects at s structure. So Bresnan has no argument that there is even a logical subject in her examples (6)-(7). Similar con­siderations hold for idioms like Time's up for .r and Fortune smiled on x (pointed out to me by Probal Dasgupta). Oblique free arguments of intransitive idioms cannot be crucial to this issue.

Moreover, for both The roof caved in (on x) and Time's up (for X), as Bresnan points out, the free argument is optional in the idiom. There­fore, the idioms may be analyzed compositionally as [[The roof caved in] on NP] and [[Time's up] for NP], with the prepositions dictating the semantic effects of adding the free argument to an already complete full phrasal idiom.

Bresnan's putative counterexamples to my claim that subject idioms do not exist (pace What's eating NP) actually prove my point. That a linguist trying to settle an important issue could only come up with

examples in which the free argument is buried in an oblique preposi­tional phrase or as the possessive in an NP increases one's confidence in the rarity of true subject idioms-idioms with fixed logical subjects and free argument positions. Bresnan gives A truck hit John vs. An idea hit John as a parting shot, but it is clear that the latter sentence contains a metaphoric use of hit that permits a wide range of subjects and ob­jects (Misfortune hit Elmer, Inspiration hit Hortense, Apathy hit the city, etc.) Moreover, it simply is not true that "It is not difficult to find other examples which support the same conclusion" -the conclusion that there are no subject/object asymmetries in idioms and ranges of predi­cates. Bresnan provides no convincing examples and even pseudo­examples are difficult to uncover, as I, a former skeptic about this sort of argument, can attest after many hours of fruitless searching."

2.1.2. Semantic Roles and the Construction or Predicate-Argument Structures

To this point I have been somewhat loose in my identification of the semantic roles associated with verbs. The assumption motivating the preceding discussion was that each semantic role assigner may. in prin­ciple. assign a unique role or a unique Set of roles. For example, al­though their logical objects are both things acted upon so that they move. throw and push need not assign precisely the same semantic role; that is. the throwee and pushee roles may be distinct. Never­theless, although semantic role assigners may assign their own unique roles, there are, apparently, linguistically significant classes of seman­tic roles. I consider terms like "agent" and "theme" as naming such semantic role classes." On this view there is no reason to exclude a given semantic role, say that role assigned by swim down the river, from being both an agent and a theme, that is, from belonging to more than

Ione semantic role class.

The grammatical machinery that makes reference to semantic role classes includes the notation of P-A structures. Consider a verb like

. put, which requires some sort of locative argument in addition to a theme, but does not specify exactly what sort of location. A variety of prepositions may be used to express put's locative argument, each as­signing this argument a different semantic role, but some locative prep­

I osition must be used.

I

, 32Chapler 2

(2.21) a. in the box b. on the table c. under the hedge d. through the window

Elmer put the porcupine e. *during the movie f. 'after the fall g. *on weekends h. 'quietly i. *aardvarks

The P-A structure for put must specify that it takes a locative argument without insisting on a particular prepositional phrase expression for the argument. Therefore, some semantic role class term, "location," is needed in P-A structures, as shown in (2.22).

(2.22) 'put' (theme, location)

From the fact that classes of semantic roles have linguistic signifi­cance, one should not conclude that all semantic roles fall into one or more linguistically significant classes. It is quite possible that the se­mantic roles assigned by some items are not classifiable. For example, the logical object of like is neither theme (thing moved) nor patient (thing affected by the action of the verb). Perhaps the only linguistically significant piece of information an English speaker knows about the role assigned by like is that it is the role assigned by like.

In addition to appearing in P-A structures, semantic role classes seem necessary to express generalizations about the organization of P-A structures within a language. In English and many other languages, it is generally true that if one of the inherent roles associated with a verb is an agent role (the role of an active, animate being who inten­tionally causes something), then this role will be assigned to the logical subject of the predicate that the verb produces. It is also generally true in these languages that theme inherent roles (roles of objects which the verb specifies to undergo a change of state or location or of which the verb indicates a location) and patient inherent roles (roles of objects that bear the brunt of the action described by the verb) are assigned by verbs-that is, they are borne by logical objects. These generalizations must be stated within the grammar of a language, where" grammar" Is taken broadly to mean the representation of a speaker's knowledge of

Levels of Syntactic Analysis 33

his language. A straightforward statement of the generalizations for English is given in (2.23).

(2.23) a. agent roles-logical subject b. theme/patient roles-logical object

The existence of generalizations like (2.23) raises two important questions. First, why should languages incorporate such generaliza­tions at all; that is, why shouldn't some verbs assign agent roles to their logical objects and have theme/patient roles assigned to their logical subjects while others conform to (2.23)? Second, assuming that a grammar incorporating generalizations like (2.23) is somehow more highly valued than a grammar without them, why do English and many other languages employ the particular generalizations in (2.23) and not, say, the reverse generalizations in (2.24)?

(2.24) a. agent roles-logical object b. theme/patient roles-logical subject

In Marantz 1982a I report the results of two experiments that I con­ducted with 3- to 5-year-old children which suggest an answer to the first question. These experiments support the hypothesis that, until about the age of 5, children's knowledge of language connects semantic roles directly to their surface structure expressions. For example, English-speaking children know that agents are placed preverbally and theme/patients postverbally. The experiments show that 3- and 4-year­old children have greater difficulty learning to use verbs whose P-A structures lead the verbs to violate these generalizations about the con­nection of semantic roles and surface positions (in active declarative sentences) than verbs whose P-A structures lead the verbs to conform to these generalizations. Consider the made-up verb moak , meaning' to pound with the elbow,' with P-A structure (2.25).

(2.25) moak 'pound with the elbow' (agent); logical subject ~ patient

Sentences like (2.26) containing moak violate English-speaking 3- and 4-year-olds' generalizations that agents come preverbally, patients postverbally.

(2.26) The book is moaking Larry.

I Levels of Syntactic Analysis 35 Chapter 2 34

(2.23) are nominative-accusative languages. The P-A structure for a My experiments demonstrate that English-speaking 3- and 4-year-olds verb meaning 'hit' in a nominative-accusative language is shown in have more difficulty learning to use verbs like moak than made-up (2.28a); that for 'hit' in an ergative language is given in (2.28b). verbs conforming to the generalizations in (2.23). When 3- and 4-year­ I olds are shown a videotape of Larry pounding a book with his elbow and are told, "The book is moaking Larry," they tend to make moak:

conform to their generalizations about the association of semantic roles and surface positions. For example, they tend to say, "Cindy is rnoak­ing the ball," when shown Cindy pounding a ball with her elbow. For the child's spontaneous use of moak to be correct in the adult language, moak would have to be given the P-A structure in (2.27), which accords with the generalizations in (2.23).

(2.27) moak 'pound with the elbow' (patient)

The performance of the children in these experiments suggests that verbs violating the generalizations in a language corresponding to gen­eralizations (2.23) in English would be regularized to conform to these generalizations through the acquisition process. Children's early lin­guistic knowledge establishes a direct connection between semantic roles and their expressions in sentences. Generalizations like (2.23) en­

sure that most verbs in a language will conform to the young child's linguistic knowledge, at least in the simple active declarative sentences he most often produces.

Returning now to the second question, given that, due to the language acquisition process, languages should change to include generalizations like (2.23), why should they come to employ the gener­alizations in (2.23) and not those in (2.24)1 Nothing in my characteriza­tion of P~A structures would demand or even suggest that predicates but not verbs should assign agent roles while verbs but not predicates assign themelpatient roles. The I-s relations subject and object are tied not to semantic roles themselves but rather to the assignment of semantic roles." Therefore, one mightexpect to find languages in which the generalizations of English (2.23) are reversed. The answer to the second question, then, would be that it is an accident that English em­ploys generalizations (2.23) in place of (2.24). In chapter 6 I claim that languages employing (2.24) do exist; they are the true ergative lan­guages. In that chapter I examine the consequences within the present theory of choosing one or the other of the sets of generalizations (2.23) and (2.24), and demonstrate that Dyirbal and Central Arctic Eskimo are true ergative languages: they choose set (2.24). Languages choosing

(2.28) a. 'hit' (patient); logical subject ~ agent b. 'hit' (agent); logical subject = patient

Note that the existence of ergative language still leaves quite a lot unexplained about generalizations like (2.23) and (2.24). For example, why should themes and patients group together; do these classes of semantic roles share some semantic property? Why should agents, pa­tients, and themes, and not other semantic role classes, like location or instrument, be canonically associated with logical subjects and objects, that is, with verbs? Also, why are there so few true ergative languages compared with nominative-accusative languages? I return 10 this last question in chapter 6.

The incorporation of generalizations like (2.23) and (2.24) within a

grammar leads one to expect the existence of two sorts of verbs asso­ciated with a single inherent semantic role. If the only role associated with an intransitive verb in a nominative-accusative language falls into the agent class, by (2.23a) one expects the verb's P-A structure to look like (2.29a). If, on the other hand, the sole role falls into the theme or patient class, generalization (2.23b) indicates that this role should be assigned by the verb, necessitating the P-A structure in (2.29b).

(2.29) a. 'intransitive-V-I' (0); logical subject = agent b. 'intransitive-V-2' (themelpatient); no logical subject

The verb in (2.29a) is a degenerate function; it takes no argument into predicates. The predicate returned by the P-A structure in (2.29a) when given no arguments assigns an agent role to its logical subject. The predicate returned by the P-A structure in (2.29b), however, assigns no semantic role; it is a degenerate predicate, requiring no arguments to make a proposition. English 511'im might be a verb of type V-I; English arrive a verb of type V-2.

(2.30) a. 'swim' (0) b. 'arrive' (theme)

1

, Chapter Z 36

Relational Grammar should be credited with first distinguishing in this manner the two classes of intransitive verbs whose P-A structures are schematized in (2.29) (see Perlmutter 1983 and the references cited there). Relational Grammarians call type V-I intransitives "unerga­tive" and type V·2 intransitives "unaccusative." Throughout the RG literature one finds convincing evidence for the validity of this distinc­tion (see, e.g., Perlmutter 1978b; Perlmutter and Postal 1978a; Harris 1982). Burzio (1981) provides considerable support from Italian for the proposed distinction between intransitive verb classes. See also Peset­sky (1982) with similar data from Russian. In section 4.1.3, I demon­strate how the characterization of the distinction between unergative and unaccusative verbs implicit in (2.30), which is essentially the char­acterization supported by Relational Grammar and Burzio (1981), cor­rectly predicts one contrasting property of these verbs within the present theory.

It is necessary to have some formal means for distinguishing verbs whose P-A structures produce predicates that assign no semantic role from verbs whose P-A structures produce role-assigning predicates. Verbs of the former sort (e.g., arrive in (2.30b» will be assigned the negative value of the feature Li.log sub]; those of the latter sort (e.g., swim in (2.30a) will be given the positive value. Technically, for a verb to be [+log sub] is for it to form a phrase that takes a logical subject at los structure. A [-log sub] verb heads an los phrase that does not take a logical subject.

Encoding the unaccusative/unergative distinction between intransi­tive vern classes in a lexical feature has clear consequences in the pres­ent theory. Lexical features like [±log sub] may be carried by affixes t

I,

and thus playa role in derivational morphology. Because [-log sub] is a lexical feature in the present theory, passivization may be analyzed as the affixation of a [-log sub] morpheme to a verb, as explained in detail in chapter 4. In chapter 41 also explore strong evidence that the feature ,[±Iog sub] cross-classifies verbs in a language. For example, Albanian ,(Hubbard 1979) has two distinct conjugational paradigms, one for [+Iog sub] verbs, one for [-log sub] verbs.

The existence of generalizations referring to semantic role classes like agent and patient is supported by the simple fact that in English, if a verb is associated with an agent and a patient as I have defined these roles, chances are overwhelming that the agent argument will behave grammatically like the agent argument associated with hit and the pa­tient argument will behave grammatically like the patient argument of

Levels or Syntactic Analysis 37

hit. Evidence for the particular formulation of these generalizations as in (2.23) and (2.24) will be drawn from the existence and behavior of unaccusative verbs and ergative languages (see section 4.1.3 and chap­ter 6).

Although classes of semantic roles do figure into generalizations like (2.23) and (2.24) and are mentioned in P-A structures, as shown in (2.21)-(2.22), it is not at all clear that any rule of grammar must refer to such classes. In the present theory, these role classes appear only in P-A structures and in generalizations like (2.23). As far as the rest of the grammar is concerned, what is important to know about semantic roles is which constituents receive semantic roles from which semantic role assigners; the identity or classification of semantic roles is irrelevant.

2.1.3. Categories at I-s Structure Although I have introduced and discussed the two basic relations that I assume exist at I-s structure-that between an argument taker and its arguments and that between a semantic role assigner and its semantic role assignee-I have yet to discuss the sorts of constituents that bear these relations. The account of constituent categories that I develop here molds together leading ideas of Montague Grammar (Dowty et at. 1981) and X-bar syntax (see. e.g., lackendoff 1977). With Montague Grammar, I assume that there are just a few basic constituent types, the remaining types being "categorially" or "functionally" defined in terms of these basic types. With X-bar syntax, I assume that the notions "head of a phrase" and "maximal projection" enter into principles of constituent structure which constrain the possible internal structure of different categories.

I assume that the two basic I-s categories are the "proposition" and the "nominal." Propositions canonically make statements and nomi­nals canonically refer to entities in the domain of discourse. As far as the syntax is concerned, however, these categories are primitive; that is, the semantic basis of these categories will not figure into any syn­tactic rule or principle. As for the nonprimitive categories, I define "predicate" as the (functional) constituent type that takes an argument into a proposition. The I-s "verb" is a function from arguments to a predicate.

I define "projection" as follows:

(2.31) A projection of a constituent X is a phrase whose category type is de­termined by X.